CC Switch 3.19.1 Changes Codex Model Routing: DeepSeek Direct, Kimi and GLM Through the Local Proxy
Crafting seamless user experiences with a passion for headless CMS, Vercel deployments, and Cloudflare optimization. I'm a Full Stack Developer with expertise in building modern web applications that are blazing fast, secure, and scalable. Let's connect and discuss how I can help you elevate your next project!
The part of a CC Switch guide that ages fastest is not the install command. It is the decision about which provider needs a local proxy. As of August 5, 2026, CC Switch 3.19.1 lets Codex connect directly to the native Responses endpoints for DeepSeek V4 Flash, Volcengine Ark Coding Plan, and Tencent Hunyuan TokenHub. Providers that expose only Chat Completions, including common Kimi, GLM, and SiliconFlow routes, still need protocol conversion.
This guide turns a detailed X article into a shorter, testable path. The first goal is not to manage eight applications. It is to connect one tool to one provider, send one acceptance request, and prove which route handled it. MCP servers, skills, sessions, memory, and failover come later.
The Three Connection Modes
| Situation | Recommended route | Reason | Acceptance check |
|---|---|---|---|
| Claude Code with Anthropic or an Anthropic-compatible endpoint | Prefer direct | Fewer moving parts | Open a new terminal and ask for only OK |
| Codex with a native Responses provider | Direct | No Responses-to-Chat conversion is needed | Confirm the model is visible and the request succeeds |
| Codex with a Chat Completions-only provider | Local proxy | CC Switch converts the protocol and maps the model | Confirm the request appears in the usage log |
CC Switch is not a model or a replacement for Claude Code or Codex. It is a cross-platform configuration manager for provider credentials, endpoints, models, routing, MCP servers, prompts, skills, sessions, usage, and backups across eight AI tools.
What Changed After the Original Guide
The source article was published on July 23, 2026 and used CC Switch 3.18.0. At that point, its Codex instructions grouped DeepSeek with providers that needed the local routing layer. Version 3.19.1 changed that path eight days later.
The official release notes state that the DeepSeek V4 Flash preset now uses a native Responses endpoint and connects directly to api.deepseek.com. Volcengine Ark Coding Plan and Tencent Hunyuan TokenHub also use direct Responses routes. The local proxy still matters, but its job is narrower. Use it when a provider lacks Responses support, when you need protocol conversion, or when logs, model mapping, circuit breaking, and failover justify another layer.
DeepSeek V4 Pro remains an exception in the 3.19.1 release. Its vendor-side Codex integration was not available at publication time. Use V4 Flash for the direct preset, or follow the current CC Switch card and documentation when V4 Pro requires routing. A provider card created before the upgrade may retain older settings, so a fresh preset is often easier to verify than a migrated card.
Where CC Switch Creates Value
The problem is configuration sprawl, not a shortage of models.
Claude Code keeps configuration under ~/.claude/. Codex uses ~/.codex/auth.json and ~/.codex/config.toml. Gemini CLI, OpenCode, OpenClaw, and Hermes use other files and workspaces. Manual editing is manageable with one tool. It becomes risky when multiple providers, credentials, model identifiers, and environment variables can all override one another.
CC Switch stores provider snapshots in SQLite and writes the selected state back to each tool's live configuration. It also centralizes MCP servers, prompt files, skills, usage tracking, sessions, backups, and an optional local proxy. The official repository uses the MIT License. The latest verified release for this article is 3.19.1.
A 20-Minute First Connection
Step 1: Back Up Before Cleaning Anything
Back up the directories that belong to the tool you already use:
~/.claude/
~/.codex/
~/.gemini/
~/.cc-switch/
Do not edit ~/.cc-switch/cc-switch.db by hand. It is a database. If your shell profile exports API variables, record them before removing anything. Those variables can override a graphical selection and are useful evidence during diagnosis.
Step 2: Install From the Official Distribution
The official website is ccswitch.io. Source code and releases live under github.com/farion1231/cc-switch. The project states that CC Switch is free and open source. A site that charges for the application or asks for your account password is not an official distribution.
On macOS, install the Homebrew cask:
brew install --cask cc-switch
Upgrade it with:
brew upgrade --cask cc-switch
On August 5, 2026, the Homebrew cask points to the signed 3.19.1 DMG. Windows users should download the .msi from the official release. Linux users can choose the .deb, .rpm, or .AppImage for their environment.
Step 3: Choose One Tool and One Provider
Start with Claude Code or Codex. Do not configure Claude Desktop, Gemini, Grok Build, OpenCode, OpenClaw, and Hermes during the same first session.
Use a built-in preset when possible. Enter only the API key, base URL, and model identifier documented by the provider. A base URL often stops at /v1 or another provider-defined root. Do not append /chat/completions unless the vendor explicitly requires it. The client may append its own endpoint and create a duplicated path that returns 404.
Step 4: Run the Smallest Acceptance Test
Enable the provider, then reopen the terminal or CLI. Claude Code can hot-switch some provider data, but a clean process removes stale state from the test.
Use one request:
Reply only with OK.
An OK response is the first gate. The second gate is operational. Confirm the active provider and model in CC Switch. If the local proxy is enabled, confirm that the request appears in the usage or request log.
Direct or Proxied Codex Routing
| Provider situation | CC Switch 3.19.1 route | Boundary |
|---|---|---|
| Official OpenAI | Direct | Preserve the official login state |
| Official DeepSeek V4 Flash endpoint | Direct | The new preset uses native Responses |
| DeepSeek V4 Pro | Follow the current card and use routing when required | Direct Codex support was not available in the 3.19.1 notes |
| Volcengine Ark Coding Plan | Direct | Do not substitute a separately billed endpoint |
| Tencent Hunyuan TokenHub | Direct | Requires a TokenHub key with Hy3 access |
| Chat-only Kimi, GLM, or SiliconFlow routes | Usually local proxy | CC Switch converts Responses and Chat formats |
Bind the local proxy to 127.0.0.1 by default. It can add protocol conversion, request logs, usage tracking, model mapping, failover, health checks, and a circuit breaker. It also adds a process, a port, and mapping state that can fail. Do not expose it on 0.0.0.0 unless you understand the network and add appropriate authentication, encryption, and firewall controls.
Map Each Error to the Correct Layer
| Symptom | Check first | Avoid as the first response |
|---|---|---|
| 401 | API key, whitespace, official versus third-party login state | Reinstalling the CLI |
404 or missing /responses |
Chat-only upstream, local route mapping, base URL | Pasting a full endpoint into the base URL |
| Model not found | Current provider catalog and preset | Guessing a model identifier |
| Old model after switching | Active card, restarted process, environment overrides | Deleting the whole user directory |
| Successful request with zero usage | Direct route, app takeover, date range | Calling all direct traffic a proxy failure |
| 401 after returning to official Codex | Version 3.19.1 or later and third-party state in auth.json |
Repeating login attempts without inspecting state |
Keep three pieces of evidence for diagnosis: the active provider card, the API format or routing screen, and the complete terminal error. Redact API keys, balances, OAuth tokens, and private referral links before sharing a screenshot.
Adopt the Workbench One Layer at a Time
After the provider works, add capabilities in this order:
- Usage and logs: prove where requests go and which model handles them.
- Prompts: manage
CLAUDE.md,AGENTS.md, andGEMINI.mdas explicit work instructions. - MCP: enable one server for one application and verify that the tool appears.
- Skills: install one repeated workflow, such as code review or weekly reporting.
- Sessions: search and resume conversations across supported tools.
- Memory and workspace files: save stable preferences, never credentials.
- Backup and sync: create a local recovery copy before WebDAV or cloud-folder sync.
The order creates attribution. When only one variable changes, a failure remains diagnosable.
Pros and Cons
| Pros | Cons |
|---|---|
| One provider and extension interface across eight AI tools | It writes live configuration, so takeover state must be understood |
| More than 50 presets reduce endpoint and field mistakes | Provider capabilities change and old tutorials age quickly |
| The proxy adds conversion, logs, and failover where needed | Another proxy means another process, port, and mapping to maintain |
| SQLite, atomic writes, and backups reduce configuration damage | Centralized backups contain more sensitive material |
| MIT-licensed and available on Windows, macOS, and Linux | Third-party privacy, pricing, and service levels remain your responsibility |
Frequently Asked Questions
Is CC Switch free?
The official repository uses the MIT License, and the project describes the application as free and open source. Model APIs, relay providers, and cloud-sync services can still charge separately.
Does DeepSeek still need the CC Switch proxy for Codex?
The official DeepSeek V4 Flash preset in 3.19.1 uses a native Responses connection. V4 Pro did not have the same direct integration when that release shipped. Older cards and aggregator endpoints may still require routing, so inspect the current API format instead of applying one rule to every DeepSeek label.
Why did a provider switch not take effect?
Confirm that you changed the correct application and enabled the intended card. Restart the terminal or IDE. Then inspect ANTHROPIC_*, OPENAI_*, GEMINI_*, and XAI_* variables that may override the live configuration.
Can the local proxy be exposed to a LAN?
It should not be the default. The route can carry credentials, prompts, and model responses. Keep it on 127.0.0.1 unless you have designed authentication, transport encryption, and firewall rules for remote access.
What should a team configure during the first week?
Connect one provider, then learn to read usage and errors. Add a second model only after that path is stable. Follow with one prompt preset, one MCP server, one skill, and a tested backup.
Sources
- Official CC Switch website
- Official farion1231/cc-switch GitHub repository
- CC Switch 3.19.1 release notes
- Official CC Switch local routing manual
- Homebrew Cask: CC Switch
Author Insight
The useful part of CC Switch is not one-click model switching. It makes configuration changes observable and recoverable. A reliable AI development environment uses a minimal acceptance request, a log, and a backup after each change. Provider routes will change again. That verification habit will age more slowly than any setup screenshot.
